CN111415361B - Method and device for estimating brain age of fetus and detecting abnormality based on deep learning - Google Patents

Method and device for estimating brain age of fetus and detecting abnormality based on deep learning Download PDF

Info

Publication number
CN111415361B
CN111415361B CN202010244415.9A CN202010244415A CN111415361B CN 111415361 B CN111415361 B CN 111415361B CN 202010244415 A CN202010244415 A CN 202010244415A CN 111415361 B CN111415361 B CN 111415361B
Authority
CN
China
Prior art keywords
age
network
brain
fetal brain
fetus
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202010244415.9A
Other languages
Chinese (zh)
Other versions
CN111415361A (en
Inventor
吴丹
施文
邹煜
颜国辉
李浩天
张祎
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University ZJU
Original Assignee
Zhejiang University ZJU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang University ZJU filed Critical Zhejiang University ZJU
Priority to CN202010244415.9A priority Critical patent/CN111415361B/en
Publication of CN111415361A publication Critical patent/CN111415361A/en
Application granted granted Critical
Publication of CN111415361B publication Critical patent/CN111415361B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/30—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30016—Brain

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • Public Health (AREA)
  • Medical Informatics (AREA)
  • Pathology (AREA)
  • Databases & Information Systems (AREA)
  • Epidemiology (AREA)
  • Primary Health Care (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Magnetic Resonance Imaging Apparatus (AREA)

Abstract

The invention discloses a method and a device for estimating brain age of a fetus brain and detecting abnormality based on deep learning. In the brain age estimation and abnormality detection method, first, a data set of a T2 weighted magnetic resonance image of a normal fetal brain is established by using a clinically conventionally acquired intrauterine T2 weighted image of a pregnant woman. Secondly, the fetal brain is segmented from the uterus by using a U-shaped network, the fetal brain age is predicted by using a depth residual error network based on an attention system, and the uncertainty of the brain age and the reliability of the fetal brain age estimation are generated. And finally, constructing a classifier according to indexes such as difference, uncertainty, reliability and the like of the actual gestational age and the predicted brain age, and judging whether the fetal brain is abnormal or not. The invention can simultaneously estimate the fetal brain age and generate indexes such as uncertainty, estimation reliability and the like for detecting the fetus with abnormal brain development, and has higher accuracy and precision and higher clinical application prospect and value.

Description

Method and device for estimating brain age of fetus and detecting abnormality based on deep learning
Technical Field
The application relates to the field of brain magnetic resonance image processing, in particular to application of deep learning, brain segmentation and brain age estimation.
Background
The age of the brain based on magnetic resonance neuroimaging is widely applied to depicting the development process of a normal brain, and the degree of deviation from the development track of the normal brain can be used as a sign and an index for measuring brain abnormality. Studies over the last decade have shown that the difference between predicted brain age and actual Physiological Age (PAD) can be a measure of the abnormal development of the brain in preterm children, the extent of brain atrophy in patients with alzheimer's disease and brain trauma, and the extent of accelerated aging in patients with schizophrenia. Fetal brain imaging has gradually become an important tool for assessing normal development of early brain, however, brain age prediction methods have not been applied to neuroimaging of fetuses. As a stable novel index, the brain age prediction can potentially evaluate the development degree of the brain of the fetus and detect abnormal development, and has very important significance and value for prenatal diagnosis.
Conventional brain age estimation methods include machine learning algorithms such as support vector regression, correlation vector regression, and gaussian process regression, as described in document 1: franke et al, 2010, Cole et al, 2015, Liem et al, 2017. In recent years, the deep learning technique is widely used in the field of medical image processing, which has more advantages and better accuracy than the conventional method, as detailed in document 2: litjens et al, 2017; shen et al, 2017. Recently, deep learning methods are also gradually used for estimation of brain age, and good prediction results are obtained, as described in detail in document 3: cole et al, 2017; jonsson et al, 2019; wang et al, 2019. On one hand, the PAD belongs to regression residual, and the change of the model can directly influence the PAD size, so that the calculation results of different models cannot be unified; on the other hand, brain age prediction models are based on data from normal populations, while abnormally developing brains contain features that are not normally involved in brain development, leading to difficulties in detection of a single marker using PAD and an interpretable reduction in PAD mechanisms.
In addition, accurate fetal brain segmentation is also beneficial to accurately measuring the fetal brain age, so that effective and accurate fetal brain segmentation and brain age regression models are developed, an effective reference index can be provided for clinical and scientific research, and the accuracy of prenatal diagnosis can be further improved.
Disclosure of Invention
In order to provide a more effective method and overcome the defects of the prior art, the invention proposes to segment the fetal brain by using deep learning, establish a model of deep ensemble learning based on an attention mechanism to predict the fetal brain age and perform abnormality detection. The method is based on big data samples of intrauterine T2 weighted images of different pregnancy collected in local hospitals, and an improved U-shaped network is constructed to segment the brain of a fetus; meanwhile, a depth residual error probability type network is established by combining an attention mechanism, and the brain age of the fetus is accurately estimated; and estimating uncertainty of the brain age prediction while network training to obtain the reliability of the predicted brain age, and detecting the fetal brain abnormality through the indexes.
In order to achieve the purpose, the invention adopts the following technical scheme to realize the purpose:
in a first aspect, the invention provides a method for estimating the brain age of a fetus based on deep ensemble learning, which comprises the following steps:
s1: acquiring a normal fetus brain T2 weighted magnetic resonance image dataset, wherein the image dataset comprises a plurality of layers of T2 weighted magnetic resonance images of normal fetuses with different gestational age, and each image is marked with a gestational age estimation value corresponding to fetus clinic;
s2: establishing a depth residual error network based on an attention mechanism, and training the depth residual error network by using the image data set in S1;
the depth residual error network based on the attention mechanism is formed by connecting a plurality of attention network modules, each attention module comprises a main branch and a mask branch, and the output of the main branch and the output of the mask branch are calculated as follows to obtain an output image of the attention network module:
xl＝Tl(xl-1)+Ml(xl-1)·Tl(xl-1)
wherein M islAnd TlRespectively representing the outputs of the mask branch and the trunk branch in the ith attention network module, xlAn output image representing the ith attention network module;
obtaining a fetal brain age prediction result in the input image by the output image of the last attention network module through an average pooling layer and a full connection layer;
s3: and taking the multilayer T2 weighted magnetic resonance image of the fetal brain to be estimated as network input, and estimating the brain age of the fetal brain by using the depth residual error network based on the attention mechanism after training in S2.
Based on the above-mentioned solution of the first aspect, the following preferred implementations can be further provided in each step.
Preferably, the depth residual error network based on the attention mechanism comprises N attention network modules in total;
in the first N-1 attention network modules, a trunk branch is formed by connecting more than two residual modules, wherein each residual module comprises continuous convolution layers of 1 × 1, 3 × 3 and 1 × 1, a batch normalization layer and a linear rectification function layer, and a mask branch is added behind the first residual module of each trunk branch; the mask branch comprises a maximum pooling layer, down-sampling is carried out next to a plurality of residual modules, then oversampling is carried out next to a symmetrical structure to form a coding-decoding structure, and sigmoid function processing is carried out on output after two convolution layers of 1 multiplied by 1 to obtain the final output of the mask branch; meanwhile, spanning connections are added in the coding-decoding structure;
in the last attention network module, the trunk branches are directly connected by a plurality of 1 × 1 convolution layers; in the mask branch, replacing a coding-decoding structure into a plurality of residual modules for connection, and outputting the residual modules activated by sigmoid; .
Preferably, in S1 and S3, the multi-layer T2 weighted magnetic resonance image of the brain is segmented from the T2 weighted magnetic resonance image in the clinical uterus of the fetus by using a trained U-Net network;
the U-Net network comprises a contraction path and an expansion path; in the contraction path, 4 times of repetition is carried out continuously, 2 convolution operations are carried out firstly and then 2 multiplied by 2 maximal pooling is carried out when each repetition is carried out, the number of channels is doubled, and the feature map output after the maximal pooling is completed in the 4 th repetition enters the expansion path after 2 convolution operations are carried out in sequence; continuously repeating for 4 times in the expansion path, wherein in each repetition, after performing 2 × 2 upsampling operation on the feature map, merging and splicing the feature map with the same image size and correspondingly output in the contraction path, and then continuously performing 2 convolution operations on the splicing result; in the convolution operations of the contraction path and the expansion path, 2 x 2 convolution operations and one dropout operation are added to the convolution operations before the last maximum pooling and the convolution operations before the first upsampling operation, and 3 x 3 convolution operations are added to the rest convolution operations; after the expansion path is repeated for the last time, converting the features into probability values through 3 multiplied by 3 convolution operation, and generating a 0-1 mask through 1 multiplied by 1 convolution operation;
and when the U-Net network is used for model training, the value of each pixel in the 0-1 mask is output by using a Softmax function, and the size of mutual information of the pixels corresponding to the true value label is used as a loss function of the network training.
Further, selecting a clinical intrauterine T2 weighted magnetic resonance image of a normal fetus, and performing quality control on data, wherein the selected data are required to ensure that the distribution of the gestational age is regular and continuous; the fetus brain in the clinical intrauterine T2 weighted magnetic resonance image is subjected to manual segmentation and labeling and then serves as a training set and a testing set of U-Net.
Furthermore, when a clinical intrauterine T2 weighted magnetic resonance image of a fetus is segmented, a 0-1 mask of the brain of the fetus is generated by using the U-Net network, and the brain image of the fetus is segmented from other tissue images; defining a 2D layer with the largest fetal brain area in the multilayer images as a middle layer, filling zero to peripheral pixels of each layer of fetal brain image to form images with the same size, and normalizing the amplitude of each individual image; successive slice images, including intermediate slices, are selected for the creation of a normal fetal brain T2 weighted magnetic resonance image dataset.
In a second aspect, the invention provides a method for detecting fetal brain abnormality based on deep ensemble learning, which comprises the following steps:
s1: acquiring a normal fetus brain T2 weighted magnetic resonance image dataset, wherein the image dataset comprises a plurality of layers of T2 weighted magnetic resonance images of normal fetuses with different gestational age, and each image is marked with a gestational age estimation value corresponding to fetus clinic;
s2: establishing a depth residual error network based on an attention mechanism, wherein the depth residual error network is formed by connecting a plurality of attention network modules, each attention module comprises a main branch and a mask branch, and the output of the main branch and the output of the mask branch are calculated as follows to obtain an output image of the attention network module:
xl＝Tl(xl-1)+Ml(xl-1)·Tl(xl-1)
wherein M islAnd TlRespectively representing the outputs of the mask branch and the trunk branch in the ith attention network module, xlAn output image representing the ith attention network module;
obtaining a fetal brain age prediction result in the input image by the output image of the last attention network module through an average pooling layer and a full connection layer;
s3: setting a loss function of the deep residual network
Figure BDA0002433598400000041
Negative log-likelihood function of probability type:
Figure BDA0002433598400000042
wherein, yiA label for the ith fetal brain age, i.e. the fetal age of the fetus; mu (x)iW) and σ2(xiW) the predicted brain age and variance of the ith fetus, respectively, c being a constant;
s4: taking a T2 weighted magnetic resonance image data set of a normal fetal brain in S1 as a training set of the depth residual error network, and performing data enhancement on the data set; based on different training set data arrangement and network parameter initialization, training the network for M times to obtain a depth residual error network with M different network parameters, and forming an integrated network;
s5: taking a multilayer T2 weighted magnetic resonance image of the fetal brain to be estimated as the input of an integration network, wherein each depth residual error network in the integration network outputs the prediction result of the fetal brain age;
s6: calculating the uncertainty of the fetal brain age according to the output result of the integrated network
Figure BDA0002433598400000043
Figure BDA0002433598400000044
Wherein the random uncertainty δa(y) and cognitive uncertainty δeThe calculation formula of (y) is as follows:
Figure BDA0002433598400000051
in the formula, mu (x, w)m) And σ2(x,wm) Respectively representing the predicted fetal brain age and the variance of the mth depth residual error network;
Figure BDA0002433598400000052
the average value of the fetal brain ages predicted by the M depth residual error networks is obtained;
s7: PAD (PAD-specific power) by using difference between predicted brain age and actual fetal age and uncertainty of fetal brain age
Figure BDA0002433598400000053
Calculating the reliability of the brain age of the fetus:
Figure BDA0002433598400000054
Figure BDA0002433598400000055
wherein: c (x, y) is the credibility of the fetal brain age calculated based on the fetal brain image x and the fetal brain age y;
removing the influence of individual gestational age by linear regression on the credibility C (x, y) of the fetal brain age to form a corrected credibility index;
obtaining an absolute value of the PAD of the fetal brain age to obtain an absolute age difference of the fetal brain age;
s8: and constructing a classification model by taking the PAD of the fetal brain age, the absolute age difference, the uncertainty and the corrected credibility as indexes, and detecting the fetal brain abnormality.
Based on the above-mentioned solution of the second aspect, the following preferred implementation modes can be further provided for each step.
Preferably, the classification model is a support vector machine.
Preferably, in the integrated network, the number M of the deep residual error networks is preferably 5.
In a third aspect, the invention provides a device for detecting fetal brain abnormality based on deep ensemble learning, which is characterized by comprising a memory and a processor;
the memory for storing a computer program;
the processor is configured to, when executing the computer program, implement the method for detecting fetal brain abnormality based on deep ensemble learning according to any one of the above aspects of the second aspect.
In a fourth aspect, the invention provides a fetal brain abnormality detection apparatus based on deep ensemble learning, which is characterized by comprising a data acquisition device, a memory and a processor;
the data acquisition device is used for acquiring clinical intrauterine T2 weighted magnetic resonance images of the fetus;
the memory is used for storing a computer program and an image acquired by the data acquisition equipment; the computer program comprises an integration network constructed in the deep ensemble learning-based fetal brain abnormality detection method according to any one of the schemes in the second aspect, and a trained U-Net network;
the processor is used for segmenting the clinical intrauterine T2 weighted magnetic resonance image of the fetus by utilizing a U-Net network when executing the computer program, segmenting the brain image of the fetus from other tissue images and obtaining a multilayer T2 weighted magnetic resonance image of the brain of the fetus; and then based on the segmentation result, implementing the fetal brain abnormality detection method based on deep ensemble learning according to any one of the above second aspect.
In a fifth aspect, the present invention provides a computer-readable storage medium, wherein the storage medium stores thereon a computer program, and when the computer program is executed by a processor, the method for detecting fetal brain abnormality based on deep ensemble learning according to any one of the above aspects of the second aspect is implemented.
Compared with the prior art, the invention has the following beneficial effects:
1) the invention segments the intrauterine fetal brain based on the deep learning and clinical routine T2 weighted magnetic resonance data set, and can accurately estimate the brain age of the fetus. The brain age estimation method applies the brain age to the neural image of the fetus, and utilizes the specially designed deep learning network to carry out corresponding parameter learning and training on clinical intrauterine magnetic resonance data, so that the brain of the fetus can be accurately segmented, and an accurate and stable estimated value of the brain age of the fetus can be obtained.
2) Compared with the prior art that the brain age model can only output a single brain age estimation value index, the method provided by the invention obtains more marking indexes including uncertainty and credibility of the estimated brain age by means of probability estimation, and can be applied to detection of fetal brain abnormality.
Therefore, the invention has important clinical application value and can provide reference for prenatal clinical examination.
Drawings
Fig. 1 is a modified version of the U-shaped depth learning network for segmenting the fetal brain from the endouterine magnetic resonance imaging.
FIG. 2 is a schematic diagram of a single depth residual network based on the attention mechanism, where the same shapes represent the same attention modules, and the preferred number of residual module channels to be composed in each attention module is written below the first residual module.
Fig. 3 is the result of predicting fetal brain age on a normal fetal test set based on deep ensemble learning.
Fig. 4 is an experimental result for distinguishing a normal fetus from various abnormal fetuses based on different markers (PAD, absolute age difference AAD, uncertainty, and gestational age reliability).
Detailed Description
The following method based on the present invention is combined with the following embodiments to show the specific technical effects thereof, so as to enable those skilled in the art to better understand the essence of the present invention.
In a preferred implementation manner of the present invention, the method for estimating the brain age of a fetus based on deep ensemble learning comprises the following steps:
the method comprises the following steps: a normal fetal brain T2 weighted magnetic resonance image dataset is established. The data set contains a plurality of multi-slice T2 weighted magnetic resonance images of normal fetal brains, each image is marked by its corresponding fetal age in fetal clinics, and a larger distribution range of fetal ages is guaranteed.
The data set is used to train the subsequent prediction network, so its sample size should meet the training requirements. The magnetic resonance scan is an intrauterine T2 weighted magnetic resonance image, which contains information about other tissues in addition to the fetal brain, and therefore requires image segmentation and extraction of the brain portion of the image. Because the sample size required by training is large, the method can additionally construct a U-Net network to segment the U-Net network.
The method for establishing the data set comprises the following steps:
1) selecting a clinical routine intrauterine T2 weighted magnetic resonance image, screening normal fetus data without intrauterine infection of a pregnant woman and obvious intracranial abnormality of a fetus, performing quality control on the data, ensuring that the selected data does not have obvious fetal movement artifacts or other serious image artifacts, and simultaneously ensuring that the head directions of the fetus acquired by magnetic resonance are consistent, and the selected data are regularly and continuously distributed as a training set and a test set of U-Net.
2) And constructing an improved U-Net network, and segmenting the multi-layer T2 weighted magnetic resonance image of the brain from the clinical intrauterine T2 weighted magnetic resonance image of the fetus by using the trained U-Net network.
The improved version of U-Net is trained based on the additional intra-uterine T2 weighted magnetic resonance image and the mask corresponding to the manually segmented fetal brain location.
As shown in fig. 1, the U-Net network of the present invention includes a contracting path (contracting path) and an expanding path (expanding path). The contraction path and the expansion path are both composed of a plurality of repeated blocks and are in mirror symmetry U-shaped structures. The network mainly comprises a convolution layer, a maximum pooling layer (down sampling), a deconvolution layer (up sampling) and a ReLU linear rectification function, and the whole network process specifically comprises the following steps:
in the contraction path, 4 times of repetition is carried out continuously, 2 convolution operations are carried out firstly and then 2 multiplied by 2 maximal pooling is carried out in each repetition, the number of channels is doubled, and the feature map output after the maximal pooling is completed in the 4 th repetition enters the expansion path after 2 convolution operations are carried out in sequence.
And in the expansion path, continuously repeating for 4 times, wherein in each repetition, the feature map is subjected to 2 × 2 upsampling operation, then is merged and spliced with the feature map with the same image size correspondingly output in the contraction path, and then continuously performing 2 convolution operations on the splicing result. It should be noted that when merging and splicing the feature maps in the contraction path and the expansion path, they should be spliced with the 2 nd convolved feature map in the corresponding block of the same layer, i.e. copy and crop (copy and crop) operations are performed. After the last iteration of the dilation path is completed, the features are transformed into probability values by a 3 × 3 convolution operation, and a 0-1 mask is generated by a 1 × 1 convolution operation.
In the traditional U-Net network, the convolution in each repeated process adopts 3 multiplied by 3 convolution operation, and the traditional U-Net network is improved in the invention: in the convolution operations of the systolic path and the expanded path, the convolution operation before the last maximum pooling and the convolution operation before the first upsampling both use a 2 × 2 convolution operation (including the ReLU linear rectification function) and add one dropout operation, while the remaining convolution layers still use a 3 × 3 convolution operation (including the ReLU linear rectification function). That is, in the 4 th iteration of the contraction path, the 2 nd convolution operation uses 2 × 2 convolution + dropout, and the convolution operation (the 2 nd convolution layer at the bottom layer) before the expansion path performs the first upsampling also uses 2 × 2 convolution + dropout, and the addition of the dropout layer can effectively prevent the occurrence of the overfitting phenomenon.
Segmenting the brain of the fetus with the selected data from other tissues, ensuring that the cuboids containing the segmented brain are minimum in volume, and defining the 2D layer with the largest fetal brain area as an intermediate layer; filling zero around each layer of fetal brain to form an image with the same size, and normalizing the amplitude of each individual image; a data set of T2 weighted magnetic resonance images of the normal fetal brain is created by taking successive slices of images including intermediate slices, represented as
Figure BDA0002433598400000081
The training set in the step 1) can be used for training the U-Net network, when a model is trained, the value of each pixel in a 0-1 mask is output by using a Softmax function, and the size of mutual information of the pixels corresponding to the true value labels is used as a loss function of network training.
The above-described U-Net, after training in the training set and validation by the test set, can be used for brain segmentation of clinically routine intrauterine T2 weighted magnetic resonance images.
When a clinical intrauterine T2 weighted magnetic resonance image of a fetus is segmented, a 0-1 mask of the brain of the fetus in the image is generated by utilizing a trained U-Net network, the brain area in the mask is 1, and the rest positions are 0. Through the mask, the brain image of the fetus can be divided from other tissue images, and the cuboid volume capable of containing the divided brain is ensured to be minimum. Defining a 2D layer with the largest fetal brain area in the multilayer images as a middle layer, filling zero to peripheral pixels of each layer of fetal brain image to form images with the same size, and normalizing the amplitude of each individual image; successive slice images including intermediate slices are selected as T2 weighted magnetic resonance images of the brain of the fetus. After the corresponding brain images of the normal fetuses with different gestational ages are obtained according to the method, the weighted magnetic resonance image data of the normal fetuses brain T2 can be establishedSet, represented as
Figure BDA0002433598400000091
Step two: establishing a depth residual error network based on an attention mechanism to predict the brain age of a fetus:
the depth residual error network based on the attention mechanism mainly comprises N attention modules, preferably three attention modules in the invention, and the structure of the depth residual error network is shown in FIG. 2. The input to the depth residual network is a weighted magnetic resonance image x of the normal fetal brain T2. Each attention module contains a trunk and a mask branch.
In the first N-1 attention network modules, a trunk branch is formed by connecting more than two residual modules, wherein each residual module (residual bounding block) includes consecutive 1 × 1, 3 × 3, and 1 × 1 convolutional layers, a batch normalization layer and a linear rectification function layer (ReLU), a mask branch is added behind the first residual module of each trunk branch, the mask branch includes a maximum pooling layer, down-sampling is performed next to the residual modules, oversampling (forming a coding-decoding structure) is performed next to a symmetric structure, sigmoid function processing is performed on output through two 1 × 1 convolutional layers, final output of the mask branch is obtained, and finally the trunk branch and the mask branch are merged to obtain an output image through the following calculation:
xl＝Tl(xl-1)+Ml(xl-1)·Tl(xl-1)
wherein M islAnd TlRepresenting the outputs of the sum mask branch and the trunk branch, x, in the ith attention module, respectivelylRepresenting the last generated image in the ith attention module. Meanwhile, a spanning connection is added in the encoding-decoding structure.
The last attention network module also comprises a trunk and a mask branch, the trunk branch is directly connected by a plurality of 1 multiplied by 1 convolution layers, the coding-decoding structure is replaced by a plurality of residual modules in the mask branch, the output is activated by sigmoid, and then the outputs of the trunk branch and the mask branch are calculated and combined as same as other attention network modules. And finally, outputting the image to obtain a result of predicting the brain age through an average pooling layer and a full connection layer.
Therefore, after the construction of the degree residual error network structure is completed, the data set of the normal fetal brain T2 weighted magnetic resonance image obtained in the step one can be used
Figure BDA0002433598400000092
The brain age estimation method is trained and can be used for subsequent brain age estimation application after training is finished. In specific application, a multilayer T2 weighted magnetic resonance image of the fetal brain to be estimated can be used as network input, and a depth residual error network based on an attention mechanism after training is utilized to output the estimated value of the brain age.
Therefore, the invention can be combined with an attention mechanism to establish a depth residual error probability type network and accurately estimate the brain age of the fetus. Moreover, compared with the prior art that the brain age model can only output a single brain age estimation value index, the method can further estimate the uncertainty of the brain age prediction while training the network to obtain more marking indexes, including uncertainty and credibility of the estimated brain age and the like, and can be applied to the detection of the fetal brain abnormality. The following discusses a specific method for further detecting fetal brain abnormality based on the above fetal brain age estimation method.
First, according to the first step and the second step, a depth residual error network based on the attention mechanism shown in fig. 2 is constructed. And performing subsequent probabilistic deep ensemble learning by using the depth residual error network structure based on the attention mechanism to obtain the uncertainty of the fetal brain age.
Assuming that the brain age of each fetus conforms to a Gaussian distribution of variance, the conditional probability of the predicted brain age of the fetus is as follows:
Figure BDA0002433598400000101
where y is fetal brain age and x and w represent T2 weighted magnetic resonance of a single fetal brain respectivelyParameters of image and training model, mu (x, w) is predicted fetal brain age, variance sigma of fetal brain age distribution2The expectation of (x, w) is a random uncertainty δa(y), expressed as follows:
Figure BDA0002433598400000102
therefore, the negative log-likelihood function is selected as the training loss function of the probabilistic deep residual network
Figure BDA0002433598400000103
Figure BDA0002433598400000104
Wherein, yiA label for the ith fetal brain age, i.e. the fetal age of the fetus; mu (x)iW) and σ2(xiW) are the predicted brain age and variance of the ith fetus, respectively, and c is a constant.
Based on the loss function, subsequent network training may be performed. Before network training, we first perform data enhancement (picture flipping, rotation, small-range translation and cropping) on the samples of the training set to amplify the sample size. The preferred learning rate for network training is 1 × 10-2The loop is iterated 100 times by using a random gradient descent algorithm.
However, unlike the method for estimating the fetal brain age, the method uses a plurality of training to obtain a plurality of networks to form an integrated network, and constructs the abnormality index in a probabilistic manner. Based on different training set data arrangement and parameter initialization, training the network for M times to obtain M networks with different parameters to obtain an integrated network, wherein the preferred training time is 5 times. The conditional probability of the brain age of the fetus is a uniform mixture of each network estimate:
Figure BDA0002433598400000105
in the formula, mu (x, w)m) And σ2(x,wm) Respectively representing the predicted fetal brain age and the variance of the mth depth residual error network;
Figure BDA0002433598400000111
and (4) predicting the average value of the brain ages of the fetuses by the M depth residual error networks.
Cognitive uncertainty δeThe estimated quantity of (y) is equal to the predicted gestational age mu (x, w) by different modelsm) Variance of (d), random uncertainty δa(y) can also be obtained simultaneously:
Figure BDA0002433598400000112
in practical use, the uncertainty delta is recognizede(y) and random uncertainty δa(y) may be output directly from the depth residual probability network.
The uncertainty of fetal brain age is formed by random uncertainty and cognitive uncertainty
Figure BDA0002433598400000113
Figure BDA0002433598400000114
Using the difference between predicted brain age and actual fetal age (PAD) and fetal uncertainty, the confidence in fetal brain age can be calculated based on the following equation:
Figure BDA0002433598400000115
Figure BDA0002433598400000116
where C (x, y) is the calculated confidence level for fetal brain age given fetal brain image x and fetal age y.
Figure BDA0002433598400000117
And
Figure BDA0002433598400000118
the predicted value and the prediction uncertainty of the fetal brain age are respectively.
In addition, since the reliability C (x, y) of the fetal brain age is related to the individual fetal age, the influence of the individual fetal age needs to be removed from the reliability C (x, y) of the fetal brain age by linear regression to form a corrected reliability index. The fitted linear formula used in making the correction is as follows:
A＝a*x+b
wherein A is credibility C (x, y) of fetal brain age, x is fetal age of fetus, a is a preset parameter (experiential value can be taken), and b is a corrected credibility index.
Therefore, the corrected reliability index b is a-a x.
In addition, another index, namely absolute age difference, is adopted in the invention, and the absolute age difference value is the absolute value of PAD of fetal brain age.
After the indexes calculated in the above steps are obtained, a classification model (for example, a support vector machine needs to be classified and trained in advance) can be constructed based on the PAD of the fetal brain age, the absolute age difference, the uncertainty, the corrected reliability and other indexes, normal and abnormal fetuses are classified, and meanwhile, a method for verifying the model effect, such as a cross verification method, is used for detecting the classification effect according to the situations of small and large fetal brain girth, malformation, ventricular abnormality and the like.
The following shows the technical effects of the method for estimating fetal brain age and detecting fetal brain abnormality based on the above method for detecting fetal brain age and method for detecting fetal brain abnormality in combination with the examples, so that those skilled in the art can better understand the essence of the present invention. The specific steps of the method are as described above and are not described in detail.
Examples
The above described methods of fetal brain age estimation and abnormality detection based on deep ensemble learning were tested in clinical routine T2 weighted magnetic resonance data of 665 normal fetuses (22-39 gestational weeks) and 46 abnormal fetuses (22-39 gestational weeks). The normal fetus is divided into a training set, a verification set and a test set of the deep integration network according to the proportion of 65% (430 cases), 15% (10) and 20% (132 cases). Abnormal fetuses included small head circumference (8 cases), enlarged ventricles (30 cases), and malformations of brain development (8 cases), with diagnoses given by clinically experienced radiologists. The specific procedure of step one is as described above, wherein the selected number of layers includes three layers below the middle layer to three layers above the middle layer, and only the specific parameters are described below. The nuclear magnetic resonance scanning is carried out by a General Electric (GE) signal HDxt 1.5T scanner used in clinical routine; t2 weighted data were acquired along the axial direction using a single shot fast spin echo sequence (SSFSE), with echo Time (TE)/repetition Time (TR) 130/2400ms, field of view (FOV) 360 × 360mm, intra-layer resolution 0.7 × 0.7mm, and slice thickness 3-4.5 mm. The gestational age of an individual is determined by the last normal menstrual period of the pregnant woman.
Establishing a sample set for training fetal brain segmentation, including 212 examples of intrauterine magnetic resonance images (22-38 gestational weeks) acquired by the same sequence parameters and corresponding masks (masks) obtained by manual segmentation, and distributing a training set, a verification set and a test set according to the ratio of 6:2: 2. This example quantifies fetal brain segmentation results and compares BET methods in the FSL kit (FMRIB Software Library v6.0) with the results shown in table 1:
TABLE 1 comparison of results obtained by segmenting fetal brains by different methods
Figure BDA0002433598400000121
Figure BDA0002433598400000131
The results of the brain age prediction in this example are shown in FIG. 2;
as shown in the attached figure 3, the method provided by the invention predicts the Mean Absolute Error (MAE) of the brain age of the normal fetus and the actual brain age (gestational age) in the test set of the normal fetus to be 0.803 weeks, R2Is 0.926.
Further, the indexes calculated by the method, including PAD, Absolute Age Difference (AAD), uncertainty (uncertaintiy) and fetal brain age reliability (confidence), are the degrees of distinction between normal fetuses and abnormal fetuses of different types, and the experimental results are shown in fig. 3, wherein the classification method is a support vector machine, and the results are presented by a working characteristic curve ROC obtained by a leave-one-out cross-validation method and an area AUC under the curve:
as can be seen from fig. 4, AAD has a high detection rate for small head circumference (AUC 0.92), an uncertainty has a high detection rate for fetal brain malformations (AUC 0.88), a confidence of gestational age has a high detection rate for small head circumference (AUC 0.92) and brain malformations (AUC 0.89), and a certain detection rate for ventricular enlargement (AUC 0.69).
In addition, in other embodiments, a device for detecting fetal brain abnormality based on deep ensemble learning may also be provided, which includes a memory and a processor;
the memory for storing a computer program;
the processor is configured to implement the method for detecting fetal brain abnormality based on deep ensemble learning as described above when executing the computer program.
It should be noted that the Memory may include a Random Access Memory (RAM) or a Non-Volatile Memory (NVM), such as at least one disk Memory. The Processor may be a general-purpose Processor, including a Central Processing Unit (CPU), a Network Processor (NP), and the like; but also Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs) or other Programmable logic devices, discrete Gate or transistor logic devices, discrete hardware components. Of course, the device should also have the necessary components to implement the program operation, such as power supply, communication bus, etc.
Additionally, in other embodiments, a deep ensemble learning based fetal brain abnormality detection kit includes a data acquisition device, a memory, and a processor;
the data acquisition device is used for acquiring clinical intrauterine T2 weighted magnetic resonance images of a fetus, and can be particularly realized by a magnetic resonance imaging system.
The memory is used for storing a computer program and an image acquired by the data acquisition equipment; the computer program comprises the integrated network constructed in the previous embodiment, and also comprises the trained U-Net network.
The processor is used for firstly segmenting a clinical intra-uterine T2 weighted magnetic resonance image of a target fetus obtained by a magnetic resonance imaging system by utilizing a U-Net network when executing the computer program, segmenting a brain image of the fetus from other tissue images and obtaining a multilayer T2 weighted magnetic resonance image of the brain of the fetus; and then based on the segmentation result, the method for detecting the fetal brain abnormality based on the deep ensemble learning is realized, and then a detection report can be further output.
In addition, in the above-mentioned complete equipment, the memory and the processor can be further integrated in the data processing device of the magnetic resonance imaging system, after the magnetic resonance imaging system acquires the corresponding data of the diagnosis object, the data can be stored in the memory, and then the processor calls the internal program to process the data, and directly outputs the result.
In addition, in other embodiments, a computer readable storage medium has stored thereon a computer program which, when executed by a processor, implements the method for fetal brain abnormality detection based on deep ensemble learning as described above.
It should be noted that the above-mentioned embodiments are only preferred embodiments of the present invention, and are not intended to limit the present invention. Various changes and modifications may be made by one of ordinary skill in the pertinent art without departing from the spirit and scope of the present invention. Therefore, the technical scheme obtained by adopting the mode of equivalent replacement or equivalent transformation is within the protection scope of the invention.

Claims (8)

1. A fetal brain age estimation method based on deep ensemble learning is characterized by comprising the following steps:
s1: acquiring a normal fetus brain T2 weighted magnetic resonance image dataset, wherein the image dataset comprises a plurality of layers of T2 weighted magnetic resonance images of normal fetuses with different gestational age, and each image is marked with a gestational age estimation value corresponding to fetus clinic;
s2: establishing a depth residual error network based on an attention mechanism, and training the depth residual error network by using the image data set in S1;
the depth residual error network based on the attention mechanism is formed by connecting a plurality of attention network modules, each attention module comprises a main branch and a mask branch, and the output of the main branch and the output of the mask branch are combined and calculated to obtain an output image of the attention network module:
xl＝Tl(xl-1)+Ml(xl-1)·Tl(xl-1)
wherein M islAnd TlRespectively representing the outputs of the mask branch and the trunk branch in the ith attention network module, xlAn output image representing the ith attention network module;
obtaining a fetal brain age prediction result in the input image by the output image of the last attention network module through an average pooling layer and a full connection layer;
s3: taking a multilayer T2 weighted magnetic resonance image of the fetal brain to be estimated as network input, and estimating the brain age of the fetal brain by using a depth residual error network based on an attention mechanism after training in S2;
the depth residual error network based on the attention mechanism comprises N attention network modules in total;
in the first N-1 attention network modules, a trunk branch is formed by connecting more than two residual modules, wherein each residual module comprises continuous convolution layers of 1 × 1, 3 × 3 and 1 × 1, a batch normalization layer and a linear rectification function layer, and a mask branch is added behind the first residual module of each trunk branch; the mask branch comprises a maximum pooling layer, down-sampling is carried out next to a plurality of residual modules, then oversampling is carried out next to a symmetrical structure to form a coding-decoding structure, and sigmoid function processing is carried out on output after two convolution layers of 1 multiplied by 1 to obtain the final output of the mask branch; meanwhile, spanning connections are added in the coding-decoding structure;
in the last attention network module, the trunk branches are directly connected by a plurality of 1 × 1 convolution layers; in the mask branch, the coding-decoding structure is replaced by a plurality of residual modules for connection, and the output is activated by sigmoid.
2. The method for estimating brain age of a fetus based on deep ensemble learning of claim 1, wherein in each of the S1 and S3, the multi-layer T2 weighted magnetic resonance image of the brain is segmented from the T2 weighted magnetic resonance image in the clinical uterus of the fetus by using a trained U-Net network;
the U-Net network comprises a contraction path and an expansion path; in the contraction path, 4 times of repetition is carried out continuously, 2 convolution operations are carried out firstly and then 2 multiplied by 2 maximal pooling is carried out when each repetition is carried out, the number of channels is doubled, and the feature map output after the maximal pooling is completed in the 4 th repetition enters the expansion path after 2 convolution operations are carried out in sequence; continuously repeating for 4 times in the expansion path, wherein in each repetition, after performing 2 × 2 upsampling operation on the feature map, merging and splicing the feature map with the same image size and correspondingly output in the contraction path, and then continuously performing 2 convolution operations on the splicing result; in the convolution operations of the contraction path and the expansion path, 2 x 2 convolution operations and one dropout operation are added to the convolution operations before the last maximum pooling and the convolution operations before the first upsampling operation, and 3 x 3 convolution operations are added to the rest convolution operations; after the expansion path is repeated for the last time, converting the features into probability values through 3 multiplied by 3 convolution operation, and generating a 0-1 mask through 1 multiplied by 1 convolution operation;
and when the U-Net network is used for model training, the value of each pixel in the 0-1 mask is output by using a Softmax function, and the size of mutual information of the pixels corresponding to the true value label is used as a loss function of the network training.
3. The method for estimating the brain age of the fetus based on the deep ensemble learning as claimed in claim 2, wherein when the clinical intrauterine T2 weighted magnetic resonance image of the fetus is segmented, the U-Net network is used to generate a 0-1 mask of the brain of the fetus, so as to segment the brain image of the fetus from other tissue images; defining a 2D layer with the largest fetal brain area in the multilayer images as a middle layer, filling zero to peripheral pixels of each layer of fetal brain image to form images with the same size, and normalizing the amplitude of each individual image; successive slice images, including intermediate slices, are selected for the creation of a normal fetal brain T2 weighted magnetic resonance image dataset.
4. A fetal brain abnormality detection method based on deep ensemble learning is characterized by comprising the following steps:
s1: acquiring a normal fetus brain T2 weighted magnetic resonance image dataset, wherein the image dataset comprises a plurality of layers of T2 weighted magnetic resonance images of normal fetuses with different gestational age, and each image is marked with a gestational age estimation value corresponding to fetus clinic;
s2: establishing a depth residual error network based on an attention mechanism, wherein the depth residual error network is formed by connecting a plurality of attention network modules, each attention module comprises a main branch and a mask branch, and the output of the main branch and the output of the mask branch are calculated as follows to obtain an output image of the attention network module:
xl＝Tl(xl-1)+Ml(xl-1)·Tl(xl-1)
wherein M islAnd TlRespectively representing the outputs of the mask branch and the trunk branch in the ith attention network module, xlAn output image representing the ith attention network module;
obtaining a fetal brain age prediction result in the input image by the output image of the last attention network module through an average pooling layer and a full connection layer;
s3: setting a loss function of the deep residual network
Figure FDA0002771623660000031
Negative log-likelihood function of probability type:
Figure FDA0002771623660000032
wherein, yiA label for the ith fetal brain age, i.e. the fetal age of the fetus; mu (x)iW) and σ2(xiW) the predicted brain age and variance of the ith fetus, respectively, c being a constant;
s4: taking a T2 weighted magnetic resonance image data set of a normal fetal brain in S1 as a training set of the depth residual error network, and performing data enhancement on the data set; based on different training set data arrangement and network parameter initialization, training the network for M times to obtain a depth residual error network with M different network parameters, and forming an integrated network;
s5: taking a multilayer T2 weighted magnetic resonance image of the fetal brain to be estimated as the input of an integration network, wherein each depth residual error network in the integration network outputs the prediction result of the fetal brain age;
s6: calculating the uncertainty of the fetal brain age according to the output result of the integrated network
Figure FDA0002771623660000033
Figure FDA0002771623660000034
Wherein the random uncertainty δa(y) and cognitive uncertainty δeThe calculation formula of (y) is as follows:
Figure FDA0002771623660000035
in the formula, mu (x, w)m) And σ2(x,wm) Respectively representing the predicted fetal brain age and the variance of the mth depth residual error network;
Figure FDA0002771623660000036
the average value of the fetal brain ages predicted by the M depth residual error networks is obtained;
s7: PAD (PAD-specific power) by using difference between predicted brain age and actual fetal age and uncertainty of fetal brain age
Figure FDA0002771623660000037
Calculating the reliability of the brain age of the fetus:
Figure FDA0002771623660000038
Figure FDA0002771623660000039
wherein: c (x, y) is the credibility of the fetal brain age calculated based on the fetal brain image x and the fetal brain age y;
removing the influence of individual gestational age by linear regression on the credibility C (x, y) of the fetal brain age to form a corrected credibility index;
obtaining an absolute value of the PAD of the fetal brain age to obtain an absolute age difference of the fetal brain age;
s8: and constructing a classification model by taking the PAD of the fetal brain age, the absolute age difference, the uncertainty and the corrected credibility as indexes, and detecting the fetal brain abnormality.
5. The method for detecting brain abnormality of fetus based on deep ensemble learning as claimed in claim 4, wherein the number M of said integration network, deep residual error network, is preferably 5.
6. A fetal brain abnormality detection device based on deep ensemble learning is characterized by comprising a memory and a processor;
the memory for storing a computer program;
the processor, when executing the computer program, is configured to implement the method for detecting fetal brain abnormality based on deep ensemble learning of claim 4 or 5.
7. A fetal brain anomaly detection device based on deep ensemble learning is characterized by comprising a data acquisition device, a memory and a processor;
the data acquisition device is used for acquiring clinical intrauterine T2 weighted magnetic resonance images of the fetus;
the memory is used for storing a computer program and an image acquired by the data acquisition equipment; the computer program comprises an integration network constructed in the deep ensemble learning based fetal brain abnormality detection method according to claim 4 or 5, and also comprises a trained U-Net network;
the processor is used for segmenting the clinical intrauterine T2 weighted magnetic resonance image of the fetus by utilizing a U-Net network when executing the computer program, segmenting the brain image of the fetus from other tissue images and obtaining a multilayer T2 weighted magnetic resonance image of the brain of the fetus; then, based on the segmentation result, the method for detecting fetal brain abnormality based on deep ensemble learning as claimed in claim 4 or 5 is implemented.
8. A computer-readable storage medium, characterized in that the storage medium has stored thereon a computer program which, when being executed by a processor, implements the method for fetal brain abnormality detection based on deep ensemble learning of claim 4 or 5.
CN202010244415.9A 2020-03-31 2020-03-31 Method and device for estimating brain age of fetus and detecting abnormality based on deep learning Active CN111415361B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010244415.9A CN111415361B (en) 2020-03-31 2020-03-31 Method and device for estimating brain age of fetus and detecting abnormality based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010244415.9A CN111415361B (en) 2020-03-31 2020-03-31 Method and device for estimating brain age of fetus and detecting abnormality based on deep learning

Publications (2)

Publication Number Publication Date
CN111415361A CN111415361A (en) 2020-07-14
CN111415361B true CN111415361B (en) 2021-01-19

Family

ID=71493560

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010244415.9A Active CN111415361B (en) 2020-03-31 2020-03-31 Method and device for estimating brain age of fetus and detecting abnormality based on deep learning

Country Status (1)

Country Link
CN (1) CN111415361B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112150428B (en) * 2020-09-18 2022-12-02 青岛大学 Medical image segmentation method based on deep learning
CN112508013B (en) * 2020-12-02 2021-07-16 哈尔滨市科佳通用机电股份有限公司 Lock catch loss fault detection method, system and device
CN112907563B (en) * 2021-03-17 2023-12-01 复旦大学附属华山医院 CTA full-automatic collateral circulation scoring method and system based on deep learning
CN114897849A (en) * 2022-05-21 2022-08-12 四川大学 Newborn auricle morphological deformity diagnosis method based on convolutional neural network

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110555828A (en) * 2019-08-08 2019-12-10 北京深睿博联科技有限责任公司 Brain age prediction method and device based on 3D convolutional neural network

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110337670B (en) * 2018-01-15 2023-09-29 深圳博脑医疗科技有限公司 Brain age testing method and brain age testing device based on magnetic resonance image
CN110097968B (en) * 2019-03-27 2020-06-30 中国科学院自动化研究所 Baby brain age prediction method and system based on resting state functional magnetic resonance image
CN110120051A (en) * 2019-05-10 2019-08-13 上海理工大学 A kind of right ventricle automatic division method based on deep learning
CN110309800B (en) * 2019-07-05 2021-07-06 中国科学技术大学 Forest fire smoke detection method and device
CN110751636B (en) * 2019-10-12 2023-04-21 天津工业大学 Fundus image retinal arteriosclerosis detection method based on improved coding and decoding network

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110555828A (en) * 2019-08-08 2019-12-10 北京深睿博联科技有限责任公司 Brain age prediction method and device based on 3D convolutional neural network

Also Published As

Publication number Publication date
CN111415361A (en) 2020-07-14

Similar Documents

Publication Publication Date Title
CN111415361B (en) Method and device for estimating brain age of fetus and detecting abnormality based on deep learning
US20210225027A1 (en) Image region localization method, image region localization apparatus, and medical image processing device
Alansary et al. PVR: patch-to-volume reconstruction for large area motion correction of fetal MRI
JP2023511300A (en) Method and system for automatically finding anatomy in medical images
CN112348785B (en) Epileptic focus positioning method and system
Tian et al. Improving in vivo human cerebral cortical surface reconstruction using data-driven super-resolution
CN111861989A (en) Method, system, terminal and storage medium for detecting midline of brain
US11600379B2 (en) Systems and methods for generating classifying and quantitative analysis reports of aneurysms from medical image data
CN109886944A (en) A kind of white matter high signal intensity detection and localization method based on multichannel chromatogram
CN111584066B (en) Brain medical image diagnosis method based on convolutional neural network and symmetric information
Hesse et al. Subcortical segmentation of the fetal brain in 3D ultrasound using deep learning
CN114565572A (en) Cerebral hemorrhage CT image classification method based on image sequence analysis
Liu et al. Automated classification and measurement of fetal ultrasound images with attention feature pyramid network
Ciceri et al. Review on deep learning fetal brain segmentation from magnetic resonance images
Liu et al. Evaluation of spatial attentive deep learning for automatic placental segmentation on longitudinal MRI
Gaudfernau et al. Analysis of the anatomical variability of Fetal brains with Corpus callosum agenesis
CN113935957A (en) Medical image comparison method and device, electronic equipment and storage medium
CN116758087B (en) Lumbar vertebra CT bone window side recess gap detection method and device
Sander et al. Reconstruction and completion of high-resolution 3D cardiac shapes using anisotropic CMRI segmentations and continuous implicit neural representations
CN111445443B (en) Early acute cerebral infarction detection method and device
Kulseng et al. Automatic placental and fetal volume estimation by a convolutional neural network
CN113160256B (en) MR image placenta segmentation method for multitasking countermeasure model
Cai et al. Pubic Symphysis-Fetal Head Segmentation Using Pure Transformer with Bi-level Routing Attention
Kalapos et al. Automated T1 and T2 mapping segmentation on cardiovascular magnetic resonance imaging using deep learning
Wu et al. Deeply supervised network for white matter hyperintensities segmentation with transfer learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant